fix: include explicit speculative batch initialization in native pin

[joelteply]

Pin the native grammar correction that preserves tagged edit arguments behind a transparent singleton allOf wrapper. The previous converter generated an empty-object-only rule for the captured EditMode union, rejecting valid search_replace arguments.

The native change is reviewed in CambrianTech/llama.cpp#66. It also explicitly initializes two existing speculative batch label_weight pointers, fixing the GCC missing-field-initializers error found by native CI without changing their zero initialization behavior.

Measured validation: the same active Release fixture fails with the old converter and passes with the fix; the Qwen XML roundtrip and existing JSON-schema suite pass. Fresh core checks and the supported GPU publisher must validate this exact pin before adoption. Recovery owns deployment identity/binding checks and observation of an actual valid edit; source tests alone do not establish coding recovery. No serving settings or model/adapters are changed here.
Further fresh CI exposed preexisting test-walk-exact build defects: unsigned window bound compared to signed loop index, and private symbols linked from a Windows DLL. The loop now uses uint32_t; the unchanged target is registered beside existing internal tests under NOT WIN32 OR NOT BUILD_SHARED_LIBS, retaining Linux and Windows static coverage. Windows static test-walk-exact compiles and links; no assertion or model scenario was removed.